Grad-CAM

Grad-CAM computes the gradient of the score for a target class with respect to the feature maps of a convolutional layer. These gradients are globally average pooled to obtain neuron importance weights. A weighted combination of the forward activation maps, followed by a ReLU, produces the final heatmap. This highlights only the regions that have a positive influence on the class of interest.

Unlike CAM, which requires a specific architecture ending with a Global Average Pooling (GAP) layer followed by a single fully-connected layer, Grad-CAM is applicable to any CNN-based architecture. It works with:

• Fully-connected layers before the output
• Models with multiple final layers
• Architectures designed for tasks beyond classification, such as image captioning and visual question answering (VQA)

Standard Grad-CAM heatmaps are coarse due to the low spatial resolution of deep convolutional features. To achieve pixel-level precision, Grad-CAM is fused with Guided Backpropagation via pointwise multiplication. This hybrid, called Guided Grad-CAM, retains the class-discriminative property of Grad-CAM while gaining the fine-grained sharpness of guided backpropagation.

By negating the gradient of a target class before the global average pooling step, Grad-CAM can generate counterfactual visualizations. These heatmaps highlight the regions that, if removed or altered, would make the network increase its confidence in the specified class. This is crucial for understanding why a model did not predict a specific outcome.

Grad-CAM extends beyond image classification. For image captioning, the gradient of a specific generated word is used to weight visual features, showing where the model 'looked' when producing that token. For Visual Question Answering (VQA), the gradient of the answer logit localizes the image evidence used to support the textual response.

The faithfulness of a Grad-CAM map is evaluated using perturbation-based metrics:

• Deletion Metric: Incrementally removes the most salient pixels and measures the drop in class confidence. A sharp drop indicates a faithful map.
• Insertion Metric: Incrementally adds the most salient pixels into a blurred baseline and measures the confidence increase.
• ROAR (RemOve And Retrain): Retrains the model on data where top-attributed pixels are removed, verifying that the map captured true model dependencies.

The foundational gradient-based visualization that computes the absolute value of the partial derivative of the class score with respect to each input pixel. Unlike Grad-CAM, saliency maps operate directly in the input space, producing fine-grained but often noisy visualizations that lack class-discriminative localization. They represent the simplest form of vanilla backpropagation for interpretability.

A visualization technique that modifies the standard backward pass through ReLU non-linearities by only propagating positive gradients corresponding to positive activations. This prevents negative gradients from flowing backward, resulting in sharper, less noisy feature visualizations than standard saliency maps. Often combined with Grad-CAM to create Guided Grad-CAM, which multiplies the fine-grained Guided Backpropagation map with the coarse Grad-CAM localization map for high-resolution, class-discriminative explanations.

A perturbation-based visual explanation method that eliminates dependency on gradients entirely. Score-CAM weights each activation map by the model's increase in confidence when that map is used as a mask on the original image. This avoids issues like gradient saturation and noisy gradients that can affect Grad-CAM. It provides more faithful and consistent localization but at a higher computational cost due to requiring multiple forward passes.

A model-agnostic perturbation method that systematically slides a gray square across the input image and measures the resulting change in prediction probability. The drop in confidence at each location produces a heatmap of importance. While computationally expensive and producing coarse maps, it serves as a ground-truth baseline for evaluating gradient-based methods like Grad-CAM because it directly measures the model's reliance on each region.

The standard evaluation protocols for attribution map quality. Deletion incrementally removes pixels from most to least important and measures the decay in prediction probability—a sharp drop indicates a good map. Insertion does the inverse, starting from a blurred baseline and adding pixels. Grad-CAM consistently achieves strong scores on these metrics, validating its ability to identify the regions the model truly relies on for classification.

An attribution method that satisfies the completeness axiom by aggregating gradients with respect to both input features and biases at every layer of the network. This ensures the sum of all attributions exactly equals the model's output. FullGrad produces sharper and more comprehensive saliency maps than Grad-CAM by capturing the contributions of bias parameters that are often ignored by other gradient-based methods.